[ Up ] [ Current Status ] [ Previous Status ]

Nonlinear Dynamics and Control of Multibody Systems, Q2 1999

Two workstations are currently being used in two experiments involving real-time control. The first experiment consists of the real time control of an underactuated vehicle (a disk on a low friction surface with two fixed thrusters). The experiment is being used to test new algorithms for controlling vehicles engaged in high-performance and aggressive maneuvers. An overhead camera and associated hardware is used to determine the state of the system. This information is relayed to one of the Intel Workstations which handles the trajectory generation and the construction of the feedback control laws required to follow the trajectory, in real-time. Commands are relayed to the vehicle via a wireless transmitter. The second experiment consists of a three degree of freedom manipulator which is interfaced to an Intel Workstation for real-time control. The purpose of this experiment is to test new algorithms being developed for aggressively controlling manipulators.

The other two workstations are being used in the Cornell Robot Soccer project. Two teams of autonomous robots are being constructed by 30 Cornell students. These two teams will engage in a simplified game of soccer. The workstations are being used to process the real-time information being provided by a vision system and to use this information for controlling and coordinating the robots in order to score goals in the opponent's area and defend goals in its own area, in real-time.

The Cornell Team has been selected to compete in a demonstration match (against a team from Carnegie-Mellon University) at the Smithsonian Institution in December, and details of the international competition may be found at http://www.mae.cornell.edu/Robocup/

Computer Modeling of Turbulent Reactive Flows

Methods for modeling turbulent, reactive flows have been pioneered by researchers in Mechanical and Aerospace Engineering at Cornell, and these are now in widespread use in industry (e.g., in the gas-turbine industry). These methods involve a particle/mesh Monte Carlo method which is computer intensive. Within the reactive flow field being considered, the fluid state is represented by a large number of particles whose properties evolve in time according to stochastic differential equations.

In the past these codes were run on workstations, and in parallel on the IBM SP2. There is a statistical error that scales as N^{-1/2}, and a bias that scales as N^{-1}, where N is the number of particles. It is highly desirable to have access to extensive computer resources so that these errors can be reduced to acceptable levels by increasing the number of particles, N.

We have been making good use of the Intel PC's to run our particle codes. The migration from the SP2 and Unix workstations for serial (single-processor) runs was very straightforward. For a typical serial run, each Intel machine is about 25% faster than a single SP2 node.

More recently, 6 of the Intel machines have been combined to form a Beowulf/Linux/MPI cluster. Preliminary parallel performance figures quoted in our previous progress report were disappointing, but installation of a new kernel that is optimized for the i686 chip set (version 2.2.1) and MPI-CH 1.1.2 (rather than version 6.1 of LAM-MPI) has resulted in significant improvements. The following table summarizes current performance of the cluster using 1, 2 and 4 nodes. For these runs the work per node is kept essentially constant, i.e., the total size of the computation increases linearly with the number of nodes, so the (relative) wall clock time would remain at unity if perfect speed-up were achieved.

The same test run on the SP2 achieves somewhat higher parallel efficiency on 4 processors, but we are quite pleased with the Beowulf results.

Last modified on: 10/12/99